Implementing integrated hierarchical sampling in R

Background

• Sampling for S-Map
• Need to locate landscape-typical sites during initial mapping phase

Integrated Hierarchical Sampling (IHS)

• builds on ‘purposive sampling’
• Zhu et al. (2008) - 10.1007/978-1-4020-8592-5 - cluster over covariates, sample where membership is high
• This time: fuzzy c-means clustering over multiple values of c
• high membership across more iterations = more regionally typical
• high membership on larger cluster numbers only = more local/niche

Evaluations

• Yang et al. (2016) - 10.2136/sssaj2015.08.0285

  • 2 desktop exercises, IHS head to head with cLHS and SRS
  • appears to outperform cLHS at watershed scale (~1:24k)
  • more stable across iterations, potentially lower n required

Software

R version

• Quarto documents - full workflow, current R stack (sf/terra, etc)
• Reverse-engineered and verified against SoLIM software outputs

CHECK_007_RvSolim.qmd
DSM_COP_20220831_Obrienl_IHS_in_R.qmd
PROCESS_003_calculate-covariates.qmd
PROCESS_004_cluster_statistics.qmd
PROCESS_005_map-clusters.qmd
PROCESS_006_choose-samples.qmd
RES_2022_IHS.qmd
SETUP_001_get-DEM.qmd
SETUP_002_clean-DEM.qmd

library(tidyverse) # wranglin'
library(terra)     # raster 
library(sf)        # vector
library(whitebox)  # geospatial analytics
library(e1071)     # clustering functions
library(fclust)    # more clustering functions

library(leaflet)   # pretty web maps
library(scico)     # pretty colours
library(flextable) # pretty tables

Local testing

Where to from here?

• Assessment
  • clearer tests vs other methods
  • incorporate wider range of covariates
  • utility for validation stage?
• Refinement
  • geocmeans package - spatially aware clustering
  • other image segmentation methods e.g. motif
• Remix

  • examine legacy samples for IHS typicality

  • Zhang et al. (2022) - cluster over different covariate mixes at one optimal c each, then IHS - optimising for multiple soil properties